Facilitating Document Reading by Linking Text and Tables
summary
An interactive PDF reader links sentences to table cells using table-structure rules, Stanford CoreNLP phrase trees and word2vec similarity, and numeric rounding heuristics. On 330 sentence–table pairs the automatic pipeline extracts 48.8% correct references with 30.5% false negatives and 20.7% false positives. The reader highlights cells or scrolls to text and, in a 14-participant study, yields a 26% increase in accuracy and about 11 seconds faster per task.
AI-written summary · about
chapters (13)
- 0:00Introduction and motivation
- 1:29Related work comparison
- 2:33System pipeline overview
- 3:17Table structure extraction
- 5:08Phrase matching with semantics
- 7:22Number and modifier handling
- 8:47Fine-tuning with table structure
- 9:51Error analysis and limitations
- 10:57Performance evaluation results
- 12:41Interactive document reader demo
- 13:24User study findings
- 15:07Future work directions
- 15:51Q&A session
related talks
10mCrowdsourcing Scholarly Discourse Annotations
12mGarden of Papers: Finding, Reading, and Organizing Research Papers in a Visual, Integrated, and F...
20mInstant Impact: Developing {docorator} to Simplify R Adoption for Teams
20mKnowledge Scout: From Talks and Slides to Smart Summaries
13mAbstractExplorer: Leveraging Structure-Mapping Theory to Enhance Comparative Close Reading at Scale
23mHidden in plain sight: Using law to summarize the law
16mSometimes you just need words
35mDocling: Get your documents ready for gen AI